Pacing Opinion Polarization via Graph Reinforcement Learning

arXiv:2602.23390v2 Announce Type: replace-cross Abstract: Opinion polarization moderation has been studied mainly as an analytical optimization problem under the Friedkin Johnson FJ model, where intervention algorithms rely on linear steady state analysis and model specific derivations. While effective in narrowly structured settings, such methods scale poorly and do not naturally extend to richer intervention regimes. This raises a central question: can polarization moderation be treated as a graph based sequential planning problem? We answer this question by proposing PACIFIER, to our knowledge the first unified graph learning framework, and in particular the first graph reinforcement learning framework, for opinion polarization moderation. PACIFIER reformulates the canonical ModerateInternal MI and ModerateExpressed ME problems as sequential decision making tasks on graphs, replacing repeated analytical recomputation with learned intervention policies. The framework has two variants: PACIFIER RL for long horizon planning and PACIFIER Greedy for efficient myopic ranking. It also extends naturally to cost aware moderation, continuous valued internal opinions, and topology altering node removal. Experiments on 15 real world polarized networks reveal a clear regime dependent picture. In analytically structured MI settings, PACIFIER remains competitive with strong analytical solvers and consistently emerges as the strongest scalable non analytical alternative. In contrast, in ME, continuous ME, and cost ME, PACIFIER achieves strong and highly consistent superiority over non PACIFIER baselines. Most importantly, PACIFIER RL becomes decisively superior in cost ME and topology altering node removal, where long horizon reasoning over future consequences is crucial. Overall, PACIFIER shifts opinion polarization moderation from model specific analytical optimization toward a unified graph learning and graph reinforcement learning paradigm.